[nabinchha]

📋 Summary

Slurm rejected every vLLM image outside 0.21.x and 0.22.x during server resolution, including newer releases and model-specific development builds that users selected explicitly. Remove that release allowlist and preserve the inspected runtime version as provenance.

Targets feat/slurm-execution.

🔗 Related Issue

Related to #1085. This removes the preflight version blocker; it does not establish Qwen3.8 model compatibility or complete the live qualification requested in that issue.

🔄 Changes

• Remove the release-line allowlist and version parsing from serving resolution. No compatibility override flag or replacement allowlist is introduced.
• Preserve the inspected runtime version unchanged, including custom/vendor version labels.
• Replace version-rejection tests with parameterized acceptance and serialization coverage for older/newer releases, prereleases, vendor builds, vLLM 0.30.0, and the exact development build reported in Slurm: support Qwen3.8-capable vLLM runtimes #1085.
• Update image-selection and getting-started documentation to explain release-independent admission and model/runtime compatibility.

Image inspection, digest verification, official requested-release matching, topology validation, and startup readiness checks remain in place. Removing the gate does not add missing model implementations to an older image or guarantee that arbitrary launch options work with every runtime.

🧪 Testing

• Full Slurm suite: pytest -q packages/data-designer-slurm/tests — 1,961 passed.
• Ruff lint and formatting checks on both changed Python files.
• Pre-commit hooks and git diff --check.
• make check-fern-docs — internal links across 101 pages and Fern validation passed (0 errors; 1 authentication-dependent redirects warning).
• Live GPU/model qualification: not run; the regression tests establish release-independent server resolution, not Qwen3.8 inference compatibility.

✅ Checklist

• Follows commit message conventions
• Commits are signed off (DCO)
• External-contributor issue requirement: N/A — maintainer-requested change.
• Public documentation updated.
• Architecture documentation: N/A — package layering and serving architecture are unchanged.

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge.

Summary

Removes the vLLM release allowlist from Slurm server resolution.

• Slurm resolves vLLM images without checking their release line.

No actionable issues found. Tests were inspected, not run.

_{Reviews (1) · Last reviewed commit: "fix(slurm): remove vLLM release allowlis..." · Reviewed by Greptile}

[github-actions]

Fern preview: https://nvidia-preview-pr-1086.docs.buildwithfern.com/nemo/datadesigner

Fern previews include the docs-website version archive with PR changes synced into latest. Notebook tutorials are rendered without execution outputs in previews.

[andreatnvidia]

LGTM. Tested Qwen3.8-Flash-Next BF16 on Slurm with vLLM 0.30.0 and 8 H100s: generation and Parquet collection succeeded. Used resume: if_possible for the pre-existing dataset-directory issue.